Chapter 6 Animation of plant development

The sequences of images used in Chapters 3 and 5 to illustrate the development of inorescences and compound leaves suggest the possibility of using computer animation to visualize plant development. From a practical perspective, computer animation oers several advantages over traditional time-lapse photography. Photography is sensitive to imperfections in the underlying experiment. A disease or even a temporary wilting of a plant may spoil months of lming. In nature, developmental processes are often masked by other phenomena. For example, the growth of leaves can be dicult to capture because of large changes in leaf positions during the day. Similarly, positions of tree branches may be aected by wind. Computer animation makes it possible to abstract from these distracting eects. Animation can be used when time-lapse photography is impractical because of the long development time of plants (e.g. years in the case of trees). In time-lapse photography, the initial position of the camera and the light conditions must be established a priori, without knowing the nal shape of the plant. In computer animation all developmental stages of the modeled plant are known in advance, allowing for optimal selection of the model view. Animation can be applied to visualize developmental mechanisms that cannot be observed directly in real plants, such as the propagation of hormones and nutrients. Motivation

134

Chapter 6. Animation of plant development

Animation oers an unprecedented means for visualizing the development of extinct plants on the basis of paleobotanical data. Discrete character of L-systems The original formalism of L-systems provides a model of development that is discrete both in time and space. Discretization in time implies that the model states are known only at specic time intervals. Discretization in space means that the range of model states is limited to a nite number. Parametric L-systems remove the latter effect by assigning continuous attributes to model components. However, the model states are still known only in discrete time intervals. This presents a problem in animation, where a smooth progression of the developing forms is desirable. This last statement requires further clarication. The very nature of animation is to produce the impression of continuous motion by displaying a sequence of still frames, captured at xed time intervals. Why is a continuous model of plant development needed if it will be used to generate a xed sequence of images in the nal account? Wouldnt it be enough to retain the standard denition of L-systems and assume time slices ne enough to produce the desired progression of forms? This approach, while feasible and useful, has three major drawbacks. A model can be constructed assuming longer or shorter time intervals, but once the choice has been made, the time step is a part of the model and cannot be changed easily. From the viewpoint of computer animation it is preferable to control the time step by a single parameter, decoupled from the underlying L-system. The continuity criteria responsible for the smooth progression of shapes during animation can be specied more easily in the continuous time domain. It would be conceptually elegant to separate model development, dened in continuous time, from model observation, taking place in discrete intervals. A developmental process is viewed as consisting of periods of continuous module expansion delimited by instantaneous module divisions. Special conditions are imposed to preserve the shape and growth rates of the organism during these qualitative changes. An analogy can be drawn to the theory of morphogenesis advanced by Thom [142], who viewed shape formation as a piecewise continuous process with singularities called catastrophes. Formally, development taking place in continuous time is captured using the formalism of timed DOL-systems. The key dierence between these L-systems and the types of L-systems considered so far lies in the denition of the derivation function. In ordinary L-systems, the derivation length is expressed as the number of derivation steps. In timed DOL-systems, the derivation length is associated with the total elapsed time since the beginning of the observation.

6.1. Timed DOL-systems

135

6.1

Timed DOL-systems

Let V be an alphabet and R the set of positive real numbers (including 0). The pair (a, ) V R is referred to as the timed letter a, and the number is called the age of a. A sequence of timed letters, x = (a1 , 1 ) . . . (an , n ) (V R) , is called a timed word over alphabet V . A timed DOL-system (tDOL-system) is a triplet G = V, , P , where V is the alphabet of the L-system, (V R)+ is a nonempty timed word over V , called the initial word, and P (V R) (V R) is a nite set of productions. Instead of writing ((a, ), (b1 , 1 ) . . . (bn , n )) P , the notation (a, ) (b1 , 1 ) . . . (bn , n ) is used. The parameter is referred to as the terminal age of the letter a, and each parameter i is the initial age assigned to the letter bi by production P . The following assumptions are made: C1. For each letter a V there exists exactly one value R such that (a, ) is the predecessor of a production in P . C2. If (a, ) is a production predecessor and (a, ) is a timed letter that occurs in the successor of some production in P , then > . According to these conditions, each letter has a uniquely dened terminal age. Furthermore, an initial age assigned to a letter by a production must be smaller than its terminal age, i.e., its lifetime ( ) must be positive. Let (a, ) (b1 , 1 ) . . . (bn , n ) be a production in P . A function D : ((V R)+ R) (V R) is called a derivation function if it has the following properties: A1. D(((a1 , 1 ) . . . (an , n )), t) = D((a1 , 1 ), t) . . . D((an , n ), t) A2. D((a, ), t) = (a, + t), if + t A3. D((a, ), t) = D((b1 , 1 ) . . . (bn , n ), t ( )), if + t > A derivation in a timed DOL-system is dened in terms of two types of time variables. Global time t is common to the entire word under consideration, while local age values i are specic to each letter ai .

Denition

Derivation

136

Chapter 6. Animation of plant development

Anabaena

Axiom A1 identies t as the variable that synchronizes the entire development, and species that the lineages of all letters can be considered independently from each other (thus, no interaction between letters is assumed). With the progress of time t, each letter grows older until its terminal age is reached (axiom A2). At this moment subdivision occurs and new letters are produced with initial age values specied by the corresponding production (axiom A3). Condition C1 guarantees that the subdivision time is dened unambiguously, hence the development proceeds in a deterministic fashion. Condition C2 guarantees that, for any value of time t, the recursive references specied by axiom A3 will eventually end. The above concepts are examined by formulating a timed DOLsystem that simulates the development of a vegetative part of the Anabaena catenula lament. Given the discrete model expressed by equation (1.1) on page 5, the corresponding tDOL-system is as follows: : p1 : p2 : p3 : p4 : (ar , 0) (ar , 1) (al , 1) (br , 1) (bl , 1) (al , 0)(br , 0) (bl , 0)(ar , 0) (ar , 0) (al , 0)

(6.1)

In accordance with the discrete model, it is assumed that all cells have the same lifetime, equal to one time unit. The derivation tree is shown in Figure 6.1. The nodes of the tree indicate production applications specied by axiom A3, and the triangular arcs represent the continuous aging processes described by axiom A2. The vertical scale indicates global time. For example, at time t = 2.75 the lament consists of three cells, bl , ar and ar , whose current age is equal to 0.75. According to the denition of time intervals corresponding to axioms A2 and A3, a production is applied after the age + t exceeds the terminal age. Consequently, at division time the old cells still exist. For example, at time t = 2.0 the lament consists of two cells, al and br , both of age = 1. The above L-system can be simplied by considering cells of type b as young forms of the cells of type a. This is suggested by Figure 6.1 where cells b simply precede cells a in time. The simplied L-system has two productions: p1 : (ar , 2) (al , 1)(ar , 0) p2 : (al , 2) (al , 0)(ar , 1) (6.2)

Model observation

The corresponding derivation tree starting from cell (ar , 1) is shown in Figure 6.2. Note the similarity to the tree from the previous example. Whether a natural developmental process or its mathematical model is considered, the choice of observation times and the act of observation should not aect the process itself. In other terms, each derived word should depend only on the total elapsed time t, and not on the partition

6.1. Timed DOL-systems

137

Figure 6.1: Derivation tree representing the continuous development of Anabaena catenula described by the L-system is equation (6.1). Sections of the triangles represent cell ages.

Figure 6.2: Derivation tree representing the continuous development of Anabaena catenula corresponding to the rules specied in equation (6.2)

138

Chapter 6. Animation of plant development

of t into intervals. The following theorem shows that timed DOLsystems satisfy this postulate. Theorem. Let G = V, , P be a tDOL-system, and x (V R)+ be a timed word over V . For any values of ta , tb 0, the following holds: D(D(x, ta ), tb ) = D(x, ta + tb ) Proof. Let us rst consider the special case where the word x consists of a single timed letter, (a0 , 0 ), and all productions in the set P take single letters into single letters. According to condition C1, there exists a unique sequence of productions from P such that: (ai , i ) (ai+1 , i+1 ), i = 0, 1, 2, . . . Let (ak , k ) be the result of the derivation of duration ta that starts from (a0 , 0 ). According to axioms A2 and A3, and assuming that ta > 0 0 , this derivation can be represented in the form D = = = = = where k = k + [ta (0 0 ) ((a0 , 0 ), ta ) D((a1 , 1 ), ta (0 0 )) D((a2 , 2 ), ta (0 0 ) (1 1 )) D((ak , k ), ta (0 0 ) (1 1 ) . . . (k1 k1 )) (ak , k ),
k 1 i=1

(i i )].

Since the sequence of recursive calls can be terminated only by an application of axiom A2, the index k and the age k satisfy the inequality k < k k . Due to condition C2, such an index k always exists and is unique. Let us now consider a derivation of duration tb > k k that starts from (ak , k ). Following the same reasoning, the result can be represented as (am , m ), where m = m + [tb (k k ) and
m1 i=k+1

(i i )]

m < m m . By substituting the value of k into the formula for m , we obtain after simple transformations m = m + [(ta + tb ) (0 0 )
m1 i=1

(i i )].

6.2. Selection of growth functions

139

Thus, the timed letter (am , m ) also results from the derivation of duration ta + tb starting with (a0 , 0 ): (am , m ) = D(D((a0 , 0 ), ta ), tb ) = D((a0 , 0 )x, ta + tb ). So far, we have considered only the case ta > 0 0 , Two other cases, namely, 0 ta 0 0 , and tb > k k tb > k k .

ta > 0 0 , 0 tb k k can be considered in a similar way. The remaining case, 0 ta 0 0 , 0 tb k k , is a straightforward consequence of condition C2. This completes the proof of the special case. In general, a derivation that starts from a word (a1 , 1 ) . . . (an , n ) can be considered as n separate derivations, each starting from a single letter. This observation applies not only to the initial word specied at time t = 0, but also to any intermediate word generated during the derivation. Consequently, any path in the derivation tree can be considered as a sequence of mappings that takes single letters into single letters. Application of the previous reasoning separately to every path concludes the proof. 2

6.2

Selection of growth functions

Timed L-systems capture qualitative changes in model topology corresponding to cell (or, in general, module) divisions, and return the age of each module as a function of the global time t. In order to complete model denition, it is also necessary to specify the shape of each module as a function of its age. Potentially, such growth functions can be estimated experimentally by observing real organisms [72, 73]. However, if detailed data is not available, growth functions can be selected from an appropriate class by choosing parameters so that the animation is smooth. This approach can be viewed as more than an ad hoc technique for constructing acceptable animated sequences. In fact, Thom presents it as a general methodology for studying morphogenesis [142, page 4]: The essence of the method to be described here consists in supposing a priori the existence of a dierential model underlying the process to be studied and, without knowing explicitly what the model is, deducing from the single assumption of its existence conclusions relating to the nature of the singularities of the process.

140

Chapter 6. Animation of plant development

A technique for computing parameters of growth functions in the case of nonbranching laments and simple branching structures is given below.

6.2.1

Development of nonbranching laments

In a simple case of geometric interpretation of timed L-systems, symbols represent cells that elongate during their lifetime and divide upon reaching terminal age. Model geometry does not change suddenly, which means that the length of each cell is a continuous function of time, and the length of a cell before subdivision is equal to the sum of the lengths of the daughter cells. Continuity requirements The above continuity requirements are formalized in the context of a tDOL-system G = V, , P as follows: R1. Let [min , ] describe the life span of a timed letter a V . The age min is the minimum of the initial age values assigned to a by the axiom or by some production in P . The terminal age is determined by the predecessor of the production acting on symbol a. The growth function g (a, ), which species the length of cell a as a function of age , must be a continuous function of parameter in the domain [min , ]. R2. For each production (a, ) (b1 , 1 ) . . . (bn , n ) in P the following equality holds:
n

g (a, ) =
i=1

g (bi , i )

(6.3)

Linear growth

In practice, requirement R1 is used to select the class of growth functions under consideration, and the equations resulting from requirement R2 are used to determine the parameters in function denitions. For example, in the case of the timed L-system specied in equation (6.2), requirement R2 takes the form g (ar , 2) = g (al , 1) + g (ar , 0) g (al , 2) = g (al , 0) + g (ar , 1). (6.4)

Let us assume that the growth functions are linear functions of time: g (al , ) = Al + Bl g (ar , ) = Ar + Br (6.5)

6.2. Selection of growth functions

141

Figure 6.3: Diagrammatic representation of the development of Anabaena catenula, with (a) linear and (b) exponential growth of cells

By substituting equations (6.5) into (6.4), we obtain 2Ar + Br = (1Al + Bl ) + (0Ar + Br ) 2Al + Bl = (0Al + Bl ) + (1Ar + Br ) or or 2Ar = Al + Bl 2Al = Ar + Br .

The desired continuity of development is provided by all solutions of this system. They can be expressed in terms of coecient c, which relates the growth rate of cells al to that of cells ar : Al = cAr Bl = Ar (2 c) Br = Ar (2c 1) Figure 6.3a illustrates the developmental process considered for c = 1. The diagram is obtained by plotting the cells in the lament as horizontal line segments on the screen. Colors indicate cell type and age. The observation time t ranges from 1 (at the top) to 7 (at the bottom), with increment t = 0.009. The slopes of the side edges of the diagram represent growth rates of the entire structure. Notice that they remain constant in the periods between cell divisions, then change. This eect is disconcerting in animation, since the rate of organism growth suddenly increases with each cell division. In order to prevent this, it is necessary to extend requirements R1 and R2 to a higher order of continuity N . Specically, equation (6.3) takes the form g
(k ) n

(a, ) =
i=1

g (k) (bi , i ) for k = 0, 1, . . . , N,

(6.6)

142

Chapter 6. Animation of plant development

Exponential growth

where g (k) (a, ) is the k th derivative of the growth function g (a, ) with respect to age . In the case of Anabaena, an attempt to achieve rst-order continuity assuming linear growth functions yields an uninteresting solution, g (al , ) = g (ar , ) 0. Thus, more complex growth functions must be considered, such as an exponential function that can be used to approximate the initial phase of sigmoidal growth. Assume that the growth function has the form g (al , ) = g (ar , ) = AeB . (6.7)

The objective is to nd values of parameters A and B that satisfy equation (6.6) for k = 0, 1. By substituting equation (6.7) into (6.6), we obtain AB k e2B = AB k eB + AB k , (6.8) which implies that zero-order continuity yields continuity of innite order in this case. Solution of equation (6.8) for any value of k yields 1+ 5 0.4812. (6.9) B = ln 2 Parameter A is a scaling factor and can be chosen arbitrarily. The corresponding diagrammatic representation of development is shown in Figure 6.3b. The side edges of the diagram, representing the growth rates of the whole structure, are smooth exponential curves.

6.2.2

Development of branching structures

The notions of tDOL-system and growth function extend in a straightforward way to L-systems with brackets. For example, the following tDOL-system describes the recursive structure of the compound leaves analyzed in Section 5.3. : (a, 0) p1 : (a, 1) (s, 0)[(b, 0)][(b, 0)](a, 0) p2 : (b, ) (a, 0) According to production p1 , apex a produces an internode s, two lateral apices b and a younger apex a. Production p2 transforms the lateral apices b into a after a delay . The daughter branches recursively repeat the development of the mother branch. Let us assume that the leaf development is rst-order continuous, yielding the following equations for k = 0, 1: g (k) (a, 1) = g (k) (s, 0) + g (k) (a, 0) g (k) (b, 0) = 0 g (k) (b, ) = g (k) (a, 0) (6.10) (6.11) (6.12)

6.2. Selection of growth functions

143

Equation (6.10) provides continuity along an existing axis upon the formation of a new internode s. Equation (6.11) species that a newly formed lateral apex b has zero length and zero growth rate. Equation (6.12) guarantees smooth transformation of apices b into apices a. Let us further assume that apices a and internodes s expand exponentially, g (a, ) = Aa eBa g (s, ) = As eBs , (6.13) (6.14)

for [0, 1]. The expansion of the lateral apices cannot be described by an exponential function, since it would not satisfy equations (6.11). Consequently, a polynomial growth function g (b, ) is chosen. Equations (6.11) and (6.12) x the functions endpoints and the tangents at the endpoints. Thus, in general, g (b, ) must be a polynomial of degree three or more, such as g (b, ) = Ab 3 + Bb 2 + Cb + Db . (6.15)

The system of equations (6.10) through (6.12) is solved using the initial size Aa and the growth rate coecient Ba as independent variables. By substituting (6.13) and (6.14) into equations (6.10) for k = 0, 1, we obtain As = Aa (eBa 1) Bs = Ba . Equations (6.11) and (6.15) yield Cb = Db = 0. Finally, substitution of (6.13) and (6.15) into (6.12) results in Ab = Bb Aa (Ba 2) 3 Aa = (3 Ba ). 2

Figure 6.4 shows a sequence of images produced by this model using values = Aa = 1 and Ba = 0.48. The branching angles are equal to 45 . The observation time t ranges from 6.9 to 7.7, with an increment of 0.2. Note the gradual formation of lateral segments. In the examples considered above, modules are represented as straight lines, with growth functions controlling line lengths. Other parameters, such as the branching angle, the diameter of segments and the size of predened surfaces, can be controlled in an analogous way. Generally, any developmental model captured by an OL-system with turtle interpretation can be converted into a tDOL-system and animated.

144

Chapter 6. Animation of plant development

Figure 6.4: Developmental sequence of a branching structure modeled using a tDOL-system